Haiku vs Sonnet (2026): I Ran Both on Real Business Jobs. Haiku Was 12 Times Cheaper, Not 20

Every "Haiku vs Sonnet" comparison I found does the same thing: it copies the price table, divides one number by the other, and tells you Haiku is 20 times cheaper. That's the list price. It's not your bill.

Your bill depends on how many tokens each model writes back, and whether the cheap one gets the job right often enough that nobody has to fix it. So I tested it on the boring work a business actually sends to an API. Here's what came back.

What do Haiku and Sonnet cost in October 2026?

Per million tokens, from Anthropic's pricing page, read October 17, 2026:

Notice what changed. A year ago the gap was 3x: Haiku 4.5 at $1 / $5 against Sonnet 4.5 at $3 / $15. If you read a Haiku vs Sonnet post written before October, it's describing a different decision. The 5.5 generation stretched the gap to 20x, and that's why the question is worth testing again.

The full rate card, including Opus and the batch discounts, is on my Anthropic API pricing page.

How I tested them

Four jobs a small business really sends to a model, each with answers I knew before running it. Both models, default settings, two full runs each, through the Claude API on October 17, 2026:

• Email routing: 30 inbound emails for a home-services company, sorted into billing, scheduling, complaint, sales lead or spam. Some are deliberately close calls (a disputed bill reads like a complaint).
• Invoice extraction: 6 messy invoices in different formats, 6 fields each, including one with a deposit already paid and one with a zero balance.
• Plan math: 8 word problems about seat minimums, overage minutes, token bills and tax-after-discount.
• Harder math: 4 multi-step problems with a trap in each (paperwork time that eats a job slot, tiered seat pricing with an annual discount, a token price that jumps over a length line).

It's a small test. 78 items per run is enough to show a cost pattern, and not enough to rank two models on intelligence. Treat it as "what happens on jobs like yours," not a benchmark.

Haiku vs Sonnet: the results

Three things stood out.

Haiku made one mistake, and it made it twice. On a roofing statement addressed "To: J. Patel, 1180 Elm St," Haiku returned the whole line as the customer name in both runs. Sonnet split it correctly both times. It's a small error, and it's exactly the kind that ends up in your CRM and nobody notices for a month.

The price gap shrank as the thinking got harder. On email routing, where the answer is one word per email, Haiku was 25 times cheaper. On the harder problems it was 9 times cheaper, because Haiku wrote about 2,000 output tokens of working per batch where Sonnet wrote about 820. Output is the expensive side of the bill, so a chattier model eats into its own discount.

Haiku was usually faster, but not always. Email routing came back in about 1.6 seconds on Haiku against 8 on Sonnet, and invoice extraction in about 8 seconds against 22. On the hardest problems Sonnet finished first, 7.3 seconds against 9.6, because it simply wrote less.

What does that mean on a monthly bill?

Take a busy small business: 10,000 inbound emails sorted, 2,000 invoices read, and 1,000 quotes or plan calculations checked in a month. Using the per-batch costs above:

That's the part the price-table posts skip. At small-business volume, both models cost less than a sandwich. The 12x gap only turns into real money at hundreds of thousands of calls, or in agent loops that resend long context on every step. Below that, choose on accuracy and speed, not the rate card. If you pay for Claude Max or Team, the monthly API credit those plans now include would cover a month like this on either model; how that credit works is on my Claude API credits page.

When should you use Haiku, and when Sonnet?

• Haiku 5.5: sorting, tagging, routing, short summaries, pulling known fields out of documents, first-draft replies a human approves, and the "worker" steps under a bigger model. Add a check for the fields that matter (my invoice miss would have been caught by a rule that rejects names containing digits).
• Sonnet 5.5: long multi-step agent work, coding, anything customer-facing that goes out without review, and judgment calls where the instructions don't cover every case. Anthropic says the same: Sonnet and Opus "remain better choices for complex agentic coding tasks."
• Long documents: Haiku 5.5's rate jumps five-fold once a single prompt passes 100,000 tokens ($0.50 / $2.50). Sonnet 5.5 keeps one rate across its full 1 million token window. Haiku is still cheaper on a 250,000-token contract ($0.125 of input against $0.50), but the gap drops from 20x to 4x, and splitting the document under the line usually beats both. The 100K line is covered in detail on my Claude Haiku 5.5 page.

The pattern that works for most businesses is both: Sonnet plans and handles the hard cases, Haiku does the volume. If you're picking one default model for a whole company, including the Claude app side, that decision lives in the best Claude model for business.

Who can set this up for a business?

The model is the cheap part. The work is deciding which jobs go to which model, writing the checks that catch the J. Patel mistakes, and wiring it into the inbox, CRM or accounting tool you already use. That's the AI implementation work I do for owner-led companies, at a published price, usually starting with one job that's eating somebody's afternoon.

If you want to try it yourself first: take 30 of last week's real emails, label them by hand, and run them through both models. If Haiku matches your labels, you just found your default.

Sources

Prices and model descriptions: Anthropic, "Pricing," platform.claude.com/docs/en/about-claude/pricing, read October 17, 2026 (Haiku 5.5, Sonnet 5.5, Haiku 4.5, Sonnet 4.6 and 4.5 rates; Haiku 5.5's 100,000-token line; full 1M context at standard pricing for 4.6 and later models except Haiku 5.5; the newer tokenizer note). Anthropic's Haiku 5.5 launch announcement, October 7, 2026 (read October 8, 2026) for the agentic-coding quote and Claude app availability. Test results: my own run through the Claude API on October 17, 2026, models claude-haiku-5-5 and claude-sonnet-5-5 at default settings, two runs each; costs are reported token counts multiplied by the list prices above. The monthly table is my arithmetic from those per-batch costs.